Capital is rotating out of GPUs into a new theme! Deep dive into Micron HBM and Marvell interconnect opportunities and risks

A few days ago, AI stocks were ugly. Many people opened their accounts and their mindset just collapsed. Then in this round of recovery, the first names to climb back were not Nvidia, which everyone watches every day, but Micron and Marvell. That matters a lot. It means Wall Street money is not giving up on AI; it just doesn’t want to blindly chase Nvidia anymore. The money is still in AI, it’s just moving to a different spot to sit.

In the past, whenever people talked about AI, the first thing they thought of was GPU and compute. But now the problem is, even if the GPU is powerful, if memory can’t keep up, the data cannot be fed in and it cannot run at full speed. Even if a data-center server farm has lots of chips, if inter-chip and rack-to-rack transmission is too slow, compute gets stuck in traffic. So why is Micron getting attention? Because it owns the blood vessel of AI memory. Why is Marvell getting attention? Because it owns the AI data-center highway.

In simple terms, the next AI opportunity is not necessarily where the story is loudest, but where the AI factory has the most critical, least replaceable bottlenecks. In today’s video, we’ll answer three questions: why Micron and Marvell were suddenly picked up by capital again; why AI memory and data-center connectivity may become the next main theme; and whether ordinary investors should still chase them now, how to chase them, and how not to rush in and end up standing at the mountain top.

Let’s start with Micron.

For a long time, many investors saw Micron as a cyclical stock. When memory prices rose, it rallied hard. When memory went down, it got hit hard too. But now the label on Micron is changing. It is no longer just a traditional memory company; it is becoming a core player in the AI memory theme.

The reason is simple. One of the biggest bottlenecks in AI servers now is HBM, high-bandwidth memory.

GPU is like the brain of AI, and HBM is the high-speed blood vessel that feeds data into that brain. If HBM is not fast enough or plentiful enough, even a very expensive, very powerful GPU can’t run at full capacity. It’s like buying a top-tier sports car with a huge engine, but the road is just a narrow alley, so the car cannot really go anywhere.

So this rebound in Micron is not the market buying a normal memory rebound; it is repricing AI memory.

Micron’s stock gain in 2025 was huge, and it stayed strong in 2026, clearly beating the market. On the earnings side, fiscal 2026 Q2 revenue reached $23.86 billion, adjusted EPS came in at $12.20, and gross margin improved materially. The company’s Q3 guidance was also strong, with revenue around $33.5 billion and gross margin around 81%.

That shows Micron is not just rising on a story; earnings are really catching up. More importantly, its hardest asset is HBM capacity.

Micron’s 2026 HBM capacity is already fully sold out and tied to long-term contracts. HBM4 has entered mass production and is shipping early, and Micron is also one of the key suppliers for NVIDIA Blackwell and the next-generation platform.

Stocks move on sentiment in the short run, earnings in the medium run, and supply-demand in the long run. If HBM remains in short supply and capacity is already locked up by major customers, Micron’s revenue visibility rises and its pricing power strengthens. That is why capital is willing to buy it again.

Now look at Marvell.

If Micron solves the blood-supply problem for the AI brain, Marvell solves the traffic problem for the AI factory.

An AI data center is not just a pile of GPUs. The hard part is how chips communicate, how servers connect, how racks transmit data, and how even different data centers coordinate. If connection speed is not fast enough, latency is not low enough, or power control is poor, the entire AI factory gets stuck in traffic. You may have bought a lot of compute, but you cannot actually run it.

That is Marvell’s value. It is not just riding the AI theme; it sits at a key point in the AI data-center upgrade.

Marvell’s stock rose sharply in 2026, and at some points even doubled. Recently it was also added to the S&P 500, and with Jensen Huang publicly backing it at Computex, market sentiment was further ignited.

But what matters most is not the short-term catalyst; it is the business structure. Marvell’s data-center business has become the clear main driver and is expected to grow about 50% in 2026. At the same time, its pipeline in custom AI chips, meaning Custom ASIC and XPU, is strong, and it has deep partnerships with several top cloud giants.

In simple terms, Microsoft, Google, Amazon, and Meta cannot keep relying on generic GPUs forever. GPUs are powerful, but they are expensive, supply is tight, and every cloud vendor has different workloads. So in the future they will care more about self-developed chips, custom chips, and data-center interconnect architecture.

Marvell sits exactly in that spot. It is not trying to replace Nvidia; it helps those cloud giants build their AI factories faster, connect them more reliably, and run them cheaper.

From a product perspective, Marvell covers key segments such as 800G, 1.6T optical modules, data-center switch chips, silicon photonics interconnect, and the NVLink Fusion ecosystem. It sounds very technical, but here is an easy way to think about it: Nvidia is the engine, Micron is the high-speed blood delivery system, and Marvell is the highway and transport hub. Even with a great engine, if the blood supply is weak or the highway is blocked, the whole system cannot run.

So why did this line suddenly get stronger? I think there are three reasons.

First, capital is rotating from chasing GPUs to hunting bottlenecks.

Over the past two years, the easiest way to play the AI rally was simply to buy Nvidia. AI training needs GPUs, cloud providers expanding compute need GPUs, and the large-model race needs GPUs. But Nvidia has already run a lot, and the market is now asking where the next wave of profits will flow.

The answer is not to randomly buy any stock with AI in its name, but to find the most constrained, most critical links in the AI factory. Right now the two most obvious bottlenecks are memory bandwidth and data-center interconnect. That is why Micron and Marvell are back on the radar.

 Second, HBM supply and demand are still tight, and capacity cannot be expanded instantly.

Many people easily mix HBM with ordinary DRAM, but they are not the same. Ordinary memory is more like a standardized product with stronger cyclical characteristics.HBM is different. It requires advanced stacking and packaging, difficult manufacturing processes, deep customer qualification, and binding to GPU platforms.

That means once demand picks up, supply will not catch up immediately. It is not like today the market is short, tomorrow the factory expands, and the day after tomorrow shipments surge. It is not that simple.

So from the standpoint of 2026, HBM is still in a state of undersupply. That gives Micron stronger pricing power and gives the market better earnings visibility.

The market loves not just a good story, but visible orders, calculable profits, and expectations that keep getting revised up. That is exactly what makes Micron attractive right now.

Third, AI data centers have entered the systems-engineering stage.

In the early AI rally, everyone mainly watched training. To train large models, the key thing was GPUs. But now AI is moving into inference and large-scale deployment.

As users increase, calls become more frequent, and models get more complex, the problems AI data centers need to solve are no longer just compute. They also need memory bandwidth, network latency, power management, custom architecture, and overall cost control.

That is why major U.S. cloud players are increasingly focused on self-developed and custom chips. Microsoft, Google, Amazon, and Meta all know they cannot rely on buying generic GPUs forever. GPU costs are too high, supply-chain pressure is too high, and each company wants its AI infrastructure better suited to its own business.

That is where Marvell’s value shows up. It can help build the highways inside the AI factory, set up the transport system, and improve data-flow efficiency.

So this rebound led by Micron and Marvell is really the result of industrial bottlenecks naturally emerging as AI develops to this stage. Once the market realizes that the next AI battle is not about who buys the most GPUs, but about who can run the AI factory fuller, faster, and cheaper, those companies get repriced.

What does this really mean for the AI industry?

I think the signal is that the AI industry is moving from GPU dominance into infrastructure revaluation.

What the big U.S. tech giants are doing now is massive AI deployment. They want AI inside search, office tools, advertising, e-commerce, cloud services, enterprise software, and even every app.

At that point, the whole AI system no longer just needs enough compute. It needs data to flow faster, memory to work more efficiently, chips to connect better, power use to stay low, and costs to stay controlled.

In other words, the AI industry is moving from single-point breakout to systems engineering.

So what changes will that bring?

The first change is that profits start flowing outward from GPUs. The earliest money was indeed concentrated in GPUs, but as AI factories get bigger, money will flow into more critical links. For example, HBM memory, custom ASICs, optical interconnect, switch chips, power delivery, and cooling.

The second change is that cloud vendors will accelerate customization.

Large U.S. cloud companies now face a real problem. AI demand is strong, but costs are also very high. GPUs are expensive, power is expensive, and data-center construction is expensive. If everything depends on buying generic GPUs, cloud vendors’ cost pressure will keep rising and their supply chain will become more passive.

So these giants will mostly self-develop or customize AI chips, and they will also optimize their own data-center interconnect architectures. That is the opportunity for companies like Marvell.

The third change is that memory’s strategic position is being redefined.

In the past, Micron was easily treated like a traditional cyclical stock. Memory prices rose, it rose. Memory prices fell, it fell. But once the AI era arrived, that logic is being rewritten.

AI models are getting larger, inference is getting more complex, and data volumes are exploding. Future AI chip competition will not just be about whose compute is stronger; it will also be about who has higher memory bandwidth, larger capacity, and lower power consumption.

That is why HBM matters. HBM is not ordinary memory; it is the key infrastructure determining whether an AI chip can really run at full speed.

So Micron is being repriced by the market now not because people suddenly like memory stocks again, but because the market has started to realize that AI factories cannot do without high-bandwidth memory.

At this point, investors naturally ask the big question: will Micron and Marvell keep going higher, or is this just a short-term move?

 Let’s start with Micron.

Micron’s setup is no longer just a simple cyclical rebound; it is a revaluation from a traditional memory stock into a core AI-memory asset.

If the market only treats Micron as a normal memory stock, its valuation logic stays cyclical. Prices rally for a while, the market gives a bit of a multiple, then once prices fall, the market slashes it back quickly.

But if the market starts treating Micron as an AI infrastructure company, its valuation logic changes.

Because HBM is not ordinary DRAM, and AI servers are not ordinary PCs or phones. What sits behind them is global cloud capex, Nvidia’s next-generation platform, and the long-term demand for large-model training and inference.

So Micron’s most important watch item is not normal memory pricing, but whether HBM can keep ramping and whether prices and margins can keep improving.

Going forward, there are three signals to watch closely. First, are HBM orders still full? Second, is gross margin still moving higher? Third, is management still raising demand expectations?

As long as those three signals stay intact, Micron’s AI-memory thesis is not over anytime soon.

So Micron’s opportunity and risk are both clear. The opportunity is undersupply in HBM and AI-memory revaluation. The risk is that once capacity is released in the future, supply-demand dynamics could change.

Now look at Marvell.

Marvell’s two core watch items are: first, the speed at which custom AI-chip revenue is realized; second, whether data-center interconnect and optical modules can keep growing as AI factories expand.

Marvell’s biggest advantage is that it is already in the supply chains of many top hyperscalers, meaning the most important cloud giants in the U.S. If those customers keep increasing their AI data-center spending, Marvell can continue to benefit.

At the same time, it is deeply tied to the NVIDIA ecosystem, and being added to the S&P 500 also brings passive-fund catalysts, so it has market attention in both the short and medium term.

But Marvell is not cheap now, and the market’s expectations are high. If orders are realized more slowly, gross margin misses expectations, or cloud capex pace becomes volatile, the stock can correct sharply.

So Marvell is better suited for trend-following and staggered positioning, not for going all-in at the hottest point of sentiment.

So what should ordinary investors do now?

The most common mistake is to believe a stock is a true theme only after it has already surged. Then once you jump in, the stock starts correcting and you immediately question life.

First, don’t just look at the percentage gain; look at the position.

If a stock is just bouncing from the bottom, and the background is earnings improvement, rising orders, and higher gross margin, then that rally has more fundamental support.

But if a stock has already accelerated for several sessions, sentiment is extremely heated, and everyone is saying it is the next big thing, then you need to be careful.

Applied to Micron and Marvell, Micron’s rebound is more driven by earnings and supply-demand. Marvell has fundamentals too, but short term it also got a boost from S&P 500 inclusion and Jensen Huang’s public support. The catalysts are strong, but they can also overheat sentiment. So Marvell is not unwatchable, but you need to guard against a sharp pullback after a fast run-up.

Also, when picking names, choose real bottlenecks, not fake AI.

The market is full of AI concepts now. Many companies hold a press conference or issue a release and immediately start talking about AI. But talking about AI does not mean making AI money.

When we pick companies, we have to look at whether there are real orders, whether there is pricing power, and whether there is a clear path to earnings realization.

If a company only has AI in its name, rides the hot topic in the news every day, but you cannot see AI revenue in the numbers, cannot see big-tech orders in the customer list, and margins are not improving, then be cautious.

Finally, positions must be layered. Never go all-in on direction.

AI may be a strong theme, but it will definitely shake out weak hands. U.S. growth stocks can rise violently, but they will not be gentle on the way down.

A more realistic approach is to split capital into three parts.

The first part, test lightly when the theme is confirmed.

The second part, add when the stock retests key moving averages or returns near the prior high platform.

The third part, leave room for earnings confirmation or opportunities after a market pullback.

Don’t see a rally and fire all your bullets at once. When the real opportunity comes, if you have no cash left, you can only watch.

If you still have cash, you have the initiative when pullbacks come. If you go all-in and the market jerks, you are only on the defensive.

The people who really make money are not the loudest, nor the fastest chasers, but the ones who can stay steady and keep executing when drawdowns hit.

I’m Xuanxuan Finance, a financial creator who doesn’t talk fluff and only helps you understand the deep logic behind U.S. stocks. If this kind of content is valuable to you, please like, subscribe, and save it. See you in the next one.